The article discusses the challenges of applying Bayesian optimisation when the hyperparameters of the Gaussian process model are unknown. The authors propose a new algorithm, HE-GP-UCB, which can handle unknown hyperparameters of any form. This algorithm is the first to have a no-regret property in such cases, working in both Bayesian and frequentist settings. The authors demonstrate the algorithm’s effectiveness through a set of toy problems, showing it can outperform the maximum likelihood estimator.
Publication date: 5 Feb 2024
Project Page: Not provided
Paper: https://arxiv.org/pdf/2402.01632
